Nome Sobrenome Time time time About Me Graduated
Nome Sobrenome. Time time time.
About Me Graduated St. Louis University with a BS in Biomedical Engineering Currently doing my masters in Computer Science at Georgia Tech Work at Idwall with R&D, focus in computer vision We’re Hiring : )
Object Tracking In Computer Vision
Detection Vs Tracking Detection Tracking Detect object independently in each Frame Predict the new location of the object in the next frame using estimated dynamics. Then update based upon measurements. Restricts Search Get improved estimates since measurement noise is tempered by priors Makes the assumption of continuous motion
What is it? Tracking Given a Model of expected motion, predict where the object will occur in the next frame, even before seeing the image Restrict Search for the object Improved estimates since measurement noise is reduced by trajectory smoothness Challenges Want to take dynamics into account Errors compound Occlusion and Disocclusion
Different Methods Kalman Filter 1 Method for tracking linear dynamic system with Gaussian noise. Particle Filter 2 Method of tracking that uses a set of particles/samples to represent the posterior distribution of some stochastic process. Deep Learning 3 Use of CNN for object detection and tracking Others 4 Mean shift Bayes Tracking SVM Template Matching Etc. .
Kalman Filter
Data Assume we have a simple state, defined by position and velocity. We don’t know the actual position and velocity, be we know the range https: //www. bzarg. com/p/how-a-kalman-filter-works-in-pictures/
Data For a Kalman Filter to work, we must assume our measured variables are random Gaussian’s.
Data Previous slide we defined position and velocity as uncorrelated. But we know that both are actually correlated.
Data The correlation is captured by the co-variance matrix.
Predict The kalman filter works by taking the current state, and predicting the next state.
Predict Since we can’t account for every variable, we add uncertainty to our measurements
Predict Every state can be moved to a range of states, which can be defined by a Gaussian with a set co-variance
Predict This produces a new Gaussian, with greater co-variance
Measure Next we take into account the measurement from the sensors. Which has a given mean and variance
Update With the sensor reading, and our estimate, we are left with two Gaussian Since we have two distributions, we can multiple them together to obtain correction
Update The best estimate comes from the overlap.
Update Which also happens to be a Gaussian Meaning we can do this process all over again.
Code Snippet
Simple Kalman Filter
Kalman Filter Tracking Pedestrian
https: //www. mathworks. com/videos/understanding-kalman-filters-part-3 -optimal-state-estimator--1490710645421. html
Particle Filter https: //www. codeproject. com/Articles/865934/Object-Tracking-Particle-Filter-with-Ease
Simple Particle Filter
Particle Filter Noisy video
Define Particles Calculate Error/Weights Resample Create Templates
Particle Filter With Changes in Appearance
How to Handle Changes in Target?
Particle Filter With Occlusion
Particle Filter with Multiple Targets
Particle Filter with Moving Camera
Why not just use deep learning? Knowing what’s happening is sometimes necessary Being able to explain when things fail is also important You can always combine with deep learning detection algorithms The more closed off the context, The easier it is to apply deep learning successfully
Perguntas!?
- Slides: 35